Farama-Foundation/Gymnasium-Robotics

What it solves

It provides a standardized collection of robotic environments for Reinforcement Learning (RL), allowing researchers and developers to train and test AI agents in simulated physical tasks without needing to build complex physics simulations from scratch.

How it works

The library uses the MuJoCo physics engine and its Python bindings to simulate robotic hardware. It implements the Gymnasium API, extending it with a GoalEnv class. This specialized API uses a dictionary-based observation space that tracks the current observation, the desired goal, and the achieved goal, which is specifically designed to support algorithms like Hindsight Experience Replay (HER).

Who it’s for

It is designed for AI researchers and developers focusing on robotic control, reinforcement learning, and multi-agent robotic systems.

Highlights

• Diverse Robot Models: Includes environments for 7-DoF robot arms (Fetch), 24-DoF anthropomorphic hands (Shadow Dexterous Hand), and 9-DoF Franka robots (Franka Kitchen).
• Complex Task Variety: Supports tasks ranging from simple reaching and pushing to complex manipulation like hammering nails, opening doors, and navigating mazes.
• Multi-Agent Support: Includes MaMuJoCo for multi-agent factorizations of robotic environments using the PettingZoo API.
• Goal-Oriented API: Features a built-in system to re-compute rewards and termination signals based on substituted goals.